[PATCH v3 0/2] Add cpu hotplug support for powerpc/perf/hv-24x7

This patchset add cpu hotplug support for hv_24x7 driver by adding
online/offline cpu hotplug function. It also add sysfs file
"cpumask" to expose current online cpu that can be used for
hv_24x7 event count.

Changelog:
v2 -> v3
- Corrected some of the typo mistakes and update commit message
  as suggested by Gautham R Shenoy.
- Added Reviewed-by tag for the first patch in the patchset.

v1 -> v2
- Changed function to pick active cpu incase of offline
  from "cpumask_any_but" to "cpumask_last", as
  cpumask_any_but function pick very next online cpu and incase where
  we are sequentially off-lining multiple cpus, "pmu_migrate_context"
  can add extra latency.
  - Suggested by: Gautham R Shenoy.

- Change documentation for cpumask and rather then hardcode the
  initialization for cpumask_attr_group, add loop to get very first
  NULL as suggested by Gautham R Shenoy.

Kajol Jain (2):
  powerpc/perf/hv-24x7: Add cpu hotplug support
  powerpc/hv-24x7: Add sysfs files inside hv-24x7 device to show cpumask

 .../sysfs-bus-event_source-devices-hv_24x7    |  7 ++
 arch/powerpc/perf/hv-24x7.c                   | 79 ++++++++++++++++++-
 include/linux/cpuhotplug.h                    |  1 +
 3 files changed, 86 insertions(+), 1 deletion(-)

-- 
2.18.2

Patch here adds cpu hotplug functions to hv_24x7 pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE" enum
is added.

The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.

The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.

Signed-off-by: Kajol Jain <redacted>
Reviewed-by: Gautham R. Shenoy <redacted>
---
 arch/powerpc/perf/hv-24x7.c | 45 +++++++++++++++++++++++++++++++++++++
 include/linux/cpuhotplug.h  |  1 +
 2 files changed, 46 insertions(+)

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index db213eb7cb02..ce4739e2b407 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -31,6 +31,8 @@ static int interface_version;
 /* Whether we have to aggregate result data for some domains. */
 static bool aggregate_result_elements;
 
+static cpumask_t hv_24x7_cpumask;
+
 static bool domain_is_valid(unsigned domain)
 {
 	switch (domain) {

@@ -1641,6 +1643,44 @@ static struct pmu h_24x7_pmu = {
 	.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
 };
 
+static int ppc_hv_24x7_cpu_online(unsigned int cpu)
+{
+	/* Make this CPU the designated target for counter collection */
+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;
+
+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);
+
+	if (target < 0 || target >= nr_cpu_ids)
+		return -1;
+
+	/* Migrate 24x7 events to the new target */
+	cpumask_set_cpu(target, &hv_24x7_cpumask);
+	perf_pmu_migrate_context(&h_24x7_pmu, cpu, target);
+
+	return 0;
+}
+
+static int hv_24x7_cpu_hotplug_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
+			  "perf/powerpc/hv_24x7:online",
+			  ppc_hv_24x7_cpu_online,
+			  ppc_hv_24x7_cpu_offline);
+}
+
 static int hv_24x7_init(void)
 {
 	int r;

@@ -1685,6 +1725,11 @@ static int hv_24x7_init(void)
 	if (r)
 		return r;
 
+	/* init cpuhotplug */
+	r = hv_24x7_cpu_hotplug_init();
+	if (r)
+		pr_err("hv_24x7: CPU hotplug init failed\n");
+
 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
 	if (r)
 		return r;

diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
index 191772d4a4d7..a2710e654b64 100644
--- a/include/linux/cpuhotplug.h
+++ b/include/linux/cpuhotplug.h

@@ -181,6 +181,7 @@ enum cpuhp_state {
 	CPUHP_AP_PERF_POWERPC_CORE_IMC_ONLINE,
 	CPUHP_AP_PERF_POWERPC_THREAD_IMC_ONLINE,
 	CPUHP_AP_PERF_POWERPC_TRACE_IMC_ONLINE,
+	CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
 	CPUHP_AP_WATCHDOG_ONLINE,
 	CPUHP_AP_WORKQUEUE_ONLINE,
 	CPUHP_AP_RCUTREE_ONLINE,

-- 
2.18.2

Patch here adds a cpumask attr to hv_24x7 pmu along with ABI documentation.

Primary use to expose the cpumask is for the perf tool which has the
capability to parse the driver sysfs folder and understand the
cpumask file. Having cpumask file will reduce the number of perf command
line parameters (will avoid "-C" option in the perf tool
command line). It can also notify the user which is
the current cpu used to retrieve the counter data.

command:# cat /sys/devices/hv_24x7/cpumask
0

Signed-off-by: Kajol Jain <redacted>
---
 .../sysfs-bus-event_source-devices-hv_24x7    |  7 ++++
 arch/powerpc/perf/hv-24x7.c                   | 36 +++++++++++++++++--
 2 files changed, 41 insertions(+), 2 deletions(-)

diff --git a/Documentation/ABI/testing/sysfs-bus-event_source-devices-hv_24x7 b/Documentation/ABI/testing/sysfs-bus-event_source-devices-hv_24x7
index e8698afcd952..f9dd3755b049 100644
--- a/Documentation/ABI/testing/sysfs-bus-event_source-devices-hv_24x7
+++ b/Documentation/ABI/testing/sysfs-bus-event_source-devices-hv_24x7

@@ -43,6 +43,13 @@ Description:	read only
 		This sysfs interface exposes the number of cores per chip
 		present in the system.
 
+What:		/sys/devices/hv_24x7/cpumask
+Date:		June 2020
+Contact:	Linux on PowerPC Developer List <linuxppc-dev@lists.ozlabs.org>
+Description:	read only
+		This sysfs file exposes the cpumask which is designated to make
+		HCALLs to retrieve hv-24x7 pmu event counter data.
+
 What:		/sys/bus/event_source/devices/hv_24x7/event_descs/<event-name>
 Date:		February 2014
 Contact:	Linux on PowerPC Developer List <linuxppc-dev@lists.ozlabs.org>

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index ce4739e2b407..3c699612d29f 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -448,6 +448,12 @@ static ssize_t device_show_string(struct device *dev,
 	return sprintf(buf, "%s\n", (char *)d->var);
 }
 
+static ssize_t cpumask_get_attr(struct device *dev,
+				struct device_attribute *attr, char *buf)
+{
+	return cpumap_print_to_pagebuf(true, buf, &hv_24x7_cpumask);
+}
+
 static ssize_t sockets_show(struct device *dev,
 			    struct device_attribute *attr, char *buf)
 {

@@ -1116,6 +1122,17 @@ static DEVICE_ATTR_RO(sockets);
 static DEVICE_ATTR_RO(chipspersocket);
 static DEVICE_ATTR_RO(coresperchip);
 
+static DEVICE_ATTR(cpumask, S_IRUGO, cpumask_get_attr, NULL);
+
+static struct attribute *cpumask_attrs[] = {
+	&dev_attr_cpumask.attr,
+	NULL,
+};
+
+static struct attribute_group cpumask_attr_group = {
+	.attrs = cpumask_attrs,
+};
+
 static struct bin_attribute *if_bin_attrs[] = {
 	&bin_attr_catalog,
 	NULL,

@@ -1143,6 +1160,11 @@ static const struct attribute_group *attr_groups[] = {
 	&event_desc_group,
 	&event_long_desc_group,
 	&if_group,
+	/*
+	 * This NULL is a placeholder for the cpumask attr which will update
+	 * onlyif cpuhotplug registration is successful
+	 */
+	NULL,
 	NULL,
 };
 

@@ -1683,7 +1705,7 @@ static int hv_24x7_cpu_hotplug_init(void)
 
 static int hv_24x7_init(void)
 {
-	int r;
+	int r, i = -1;
 	unsigned long hret;
 	struct hv_perf_caps caps;
 

@@ -1727,8 +1749,18 @@ static int hv_24x7_init(void)
 
 	/* init cpuhotplug */
 	r = hv_24x7_cpu_hotplug_init();
-	if (r)
+	if (r) {
 		pr_err("hv_24x7: CPU hotplug init failed\n");
+	} else {
+		/*
+		 * Cpu hotplug init is successful, add the
+		 * cpumask file as part of pmu attr group and
+		 * assign it to very first NULL location.
+		 */
+		while (attr_groups[++i])
+			/* nothing */;
+		attr_groups[i] = &cpumask_attr_group;
+	}
 
 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
 	if (r)

-- 
2.18.2

On Fri, Jun 26, 2020 at 03:58:24PM +0530, Kajol Jain wrote:

Patch here adds a cpumask attr to hv_24x7 pmu along with ABI documentation.

Primary use to expose the cpumask is for the perf tool which has the
capability to parse the driver sysfs folder and understand the
cpumask file. Having cpumask file will reduce the number of perf command
line parameters (will avoid "-C" option in the perf tool
command line). It can also notify the user which is
the current cpu used to retrieve the counter data.

command:# cat /sys/devices/hv_24x7/cpumask
0

Signed-off-by: Kajol Jain <redacted>

This patch looks good to me.

Reviewed-by: Gautham R. Shenoy <redacted>

Kajol Jain [off-list ref] writes:

The comment implies every newly onlined CPU will become the target, but
actually it's only the first onlined CPU.

So I think the comment needs updating, or you could just drop the
comment, I think the code is fairly clear by itself.

+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;

No need to initialise target, you assign to it unconditionally below.

+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);

Any reason to use cpumask_last() vs cpumask_first(), or a randomly
chosen CPU?

The hotplug initialisation shouldn't fail unless something is badly
wrong. I think you should just fail initialisation of the entire PMU if
that happens, which will make the error handling in the next patch much
simpler.

cheers

 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
 	if (r)
 		return r;

On 7/6/20 8:43 AM, Michael Ellerman wrote:

Kajol Jain [off-list ref] writes:

quoted

Patch here adds cpu hotplug functions to hv_24x7 pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE" enum
is added.

The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.

The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.

Signed-off-by: Kajol Jain <redacted>
Reviewed-by: Gautham R. Shenoy <redacted>
---
  arch/powerpc/perf/hv-24x7.c | 45 +++++++++++++++++++++++++++++++++++++
  include/linux/cpuhotplug.h  |  1 +
  2 files changed, 46 insertions(+)

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index db213eb7cb02..ce4739e2b407 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -31,6 +31,8 @@ static int interface_version;
  /* Whether we have to aggregate result data for some domains. */
  static bool aggregate_result_elements;
  
+static cpumask_t hv_24x7_cpumask;
+
  static bool domain_is_valid(unsigned domain)
  {
  	switch (domain) {

@@ -1641,6 +1643,44 @@ static struct pmu h_24x7_pmu = {
  	.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
  };
  
+static int ppc_hv_24x7_cpu_online(unsigned int cpu)
+{
+	/* Make this CPU the designated target for counter collection */

The comment implies every newly onlined CPU will become the target, but
actually it's only the first onlined CPU.

So I think the comment needs updating, or you could just drop the
comment, I think the code is fairly clear by itself.

quoted

+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;

No need to initialise target, you assign to it unconditionally below.

quoted

+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);

Any reason to use cpumask_last() vs cpumask_first(), or a randomly
chosen CPU?

quoted

+	if (target < 0 || target >= nr_cpu_ids)
+		return -1;
+
+	/* Migrate 24x7 events to the new target */
+	cpumask_set_cpu(target, &hv_24x7_cpumask);
+	perf_pmu_migrate_context(&h_24x7_pmu, cpu, target);
+
+	return 0;
+}
+
+static int hv_24x7_cpu_hotplug_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
+			  "perf/powerpc/hv_24x7:online",
+			  ppc_hv_24x7_cpu_online,
+			  ppc_hv_24x7_cpu_offline);
+}
+
  static int hv_24x7_init(void)
  {
  	int r;

@@ -1685,6 +1725,11 @@ static int hv_24x7_init(void)
  	if (r)
  		return r;
  
+	/* init cpuhotplug */
+	r = hv_24x7_cpu_hotplug_init();
+	if (r)
+		pr_err("hv_24x7: CPU hotplug init failed\n");
+

The hotplug initialisation shouldn't fail unless something is badly
wrong. I think you should just fail initialisation of the entire PMU if
that happens, which will make the error handling in the next patch much
simpler.

We  did fail the PMU registration on failure of the hotplug
code (and yes error handling is much simpler), but on internal 
review/discussion,
what came up was that, hv_24x7 PMU will still be usable without
the hotplug code (with "-C" option to perf tool command line).

Maddy

cheers

quoted

  	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
  	if (r)
  		return r;

Madhavan Srinivasan [off-list ref] writes:

On 7/6/20 8:43 AM, Michael Ellerman wrote:

quoted

Kajol Jain [off-list ref] writes:

quoted

Patch here adds cpu hotplug functions to hv_24x7 pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE" enum
is added.

The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.

The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.

Signed-off-by: Kajol Jain <redacted>
Reviewed-by: Gautham R. Shenoy <redacted>
---
  arch/powerpc/perf/hv-24x7.c | 45 +++++++++++++++++++++++++++++++++++++
  include/linux/cpuhotplug.h  |  1 +
  2 files changed, 46 insertions(+)

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index db213eb7cb02..ce4739e2b407 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -31,6 +31,8 @@ static int interface_version;
  /* Whether we have to aggregate result data for some domains. */
  static bool aggregate_result_elements;
  
+static cpumask_t hv_24x7_cpumask;
+
  static bool domain_is_valid(unsigned domain)
  {
  	switch (domain) {

@@ -1641,6 +1643,44 @@ static struct pmu h_24x7_pmu = {
  	.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
  };
  
+static int ppc_hv_24x7_cpu_online(unsigned int cpu)
+{
+	/* Make this CPU the designated target for counter collection */

The comment implies every newly onlined CPU will become the target, but
actually it's only the first onlined CPU.

So I think the comment needs updating, or you could just drop the
comment, I think the code is fairly clear by itself.

quoted

+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;

No need to initialise target, you assign to it unconditionally below.

quoted

+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);

Any reason to use cpumask_last() vs cpumask_first(), or a randomly
chosen CPU?

quoted

+	if (target < 0 || target >= nr_cpu_ids)
+		return -1;
+
+	/* Migrate 24x7 events to the new target */
+	cpumask_set_cpu(target, &hv_24x7_cpumask);
+	perf_pmu_migrate_context(&h_24x7_pmu, cpu, target);
+
+	return 0;
+}
+
+static int hv_24x7_cpu_hotplug_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
+			  "perf/powerpc/hv_24x7:online",
+			  ppc_hv_24x7_cpu_online,
+			  ppc_hv_24x7_cpu_offline);
+}
+
  static int hv_24x7_init(void)
  {
  	int r;

@@ -1685,6 +1725,11 @@ static int hv_24x7_init(void)
  	if (r)
  		return r;
  
+	/* init cpuhotplug */
+	r = hv_24x7_cpu_hotplug_init();
+	if (r)
+		pr_err("hv_24x7: CPU hotplug init failed\n");
+

The hotplug initialisation shouldn't fail unless something is badly
wrong. I think you should just fail initialisation of the entire PMU if
that happens, which will make the error handling in the next patch much
simpler.

We  did fail the PMU registration on failure of the hotplug
code (and yes error handling is much simpler), but on internal 
review/discussion,
what came up was that, hv_24x7 PMU will still be usable without
the hotplug code (with "-C" option to perf tool command line).

In theory yes.

But in reality no one will ever test that case, so the code will easily
bit rot.

Even if it doesn't bit rot, you've now created another state the system
can legally be in (hotplug init failed but PMU still probed), which you
have to test, document & support.

If the hotplug init fails then something is badly wrong, the best thing
we can do is bail on the PMU initialisation and hope the rest of the
system boots OK.

cheers

On 7/7/20 10:26 AM, Michael Ellerman wrote:

Madhavan Srinivasan [off-list ref] writes:

quoted

On 7/6/20 8:43 AM, Michael Ellerman wrote:

quoted

Kajol Jain [off-list ref] writes:

quoted

Patch here adds cpu hotplug functions to hv_24x7 pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE" enum
is added.

The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.

The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.

Signed-off-by: Kajol Jain <redacted>
Reviewed-by: Gautham R. Shenoy <redacted>
---
   arch/powerpc/perf/hv-24x7.c | 45 +++++++++++++++++++++++++++++++++++++
   include/linux/cpuhotplug.h  |  1 +
   2 files changed, 46 insertions(+)

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index db213eb7cb02..ce4739e2b407 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -31,6 +31,8 @@ static int interface_version;
   /* Whether we have to aggregate result data for some domains. */
   static bool aggregate_result_elements;
   
+static cpumask_t hv_24x7_cpumask;
+
   static bool domain_is_valid(unsigned domain)
   {
   	switch (domain) {

@@ -1641,6 +1643,44 @@ static struct pmu h_24x7_pmu = {
   	.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
   };
   
+static int ppc_hv_24x7_cpu_online(unsigned int cpu)
+{
+	/* Make this CPU the designated target for counter collection */

The comment implies every newly onlined CPU will become the target, but
actually it's only the first onlined CPU.

So I think the comment needs updating, or you could just drop the
comment, I think the code is fairly clear by itself.

quoted

+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;

No need to initialise target, you assign to it unconditionally below.

quoted

+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);

Any reason to use cpumask_last() vs cpumask_first(), or a randomly
chosen CPU?

quoted

+	if (target < 0 || target >= nr_cpu_ids)
+		return -1;
+
+	/* Migrate 24x7 events to the new target */
+	cpumask_set_cpu(target, &hv_24x7_cpumask);
+	perf_pmu_migrate_context(&h_24x7_pmu, cpu, target);
+
+	return 0;
+}
+
+static int hv_24x7_cpu_hotplug_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
+			  "perf/powerpc/hv_24x7:online",
+			  ppc_hv_24x7_cpu_online,
+			  ppc_hv_24x7_cpu_offline);
+}
+
   static int hv_24x7_init(void)
   {
   	int r;

@@ -1685,6 +1725,11 @@ static int hv_24x7_init(void)
   	if (r)
   		return r;
   
+	/* init cpuhotplug */
+	r = hv_24x7_cpu_hotplug_init();
+	if (r)
+		pr_err("hv_24x7: CPU hotplug init failed\n");
+

The hotplug initialisation shouldn't fail unless something is badly
wrong. I think you should just fail initialisation of the entire PMU if
that happens, which will make the error handling in the next patch much
simpler.

We  did fail the PMU registration on failure of the hotplug
code (and yes error handling is much simpler), but on internal
review/discussion,
what came up was that, hv_24x7 PMU will still be usable without
the hotplug code (with "-C" option to perf tool command line).

In theory yes.

But in reality no one will ever test that case, so the code will easily
bit rot.

Even if it doesn't bit rot, you've now created another state the system
can legally be in (hotplug init failed but PMU still probed), which you
have to test, document & support.

If the hotplug init fails then something is badly wrong, the best thing
we can do is bail on the PMU initialisation and hope the rest of the
system boots OK.

Yep agreed. Thanks for the comments mpe

Maddy

cheers

On 7/6/20 8:43 AM, Michael Ellerman wrote:

Kajol Jain [off-list ref] writes:

quoted

Patch here adds cpu hotplug functions to hv_24x7 pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE" enum
is added.

The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.

The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.

Signed-off-by: Kajol Jain <redacted>
Reviewed-by: Gautham R. Shenoy <redacted>
---
 arch/powerpc/perf/hv-24x7.c | 45 +++++++++++++++++++++++++++++++++++++
 include/linux/cpuhotplug.h  |  1 +
 2 files changed, 46 insertions(+)

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
index db213eb7cb02..ce4739e2b407 100644
--- a/arch/powerpc/perf/hv-24x7.c
+++ b/arch/powerpc/perf/hv-24x7.c

@@ -31,6 +31,8 @@ static int interface_version;
 /* Whether we have to aggregate result data for some domains. */
 static bool aggregate_result_elements;
 
+static cpumask_t hv_24x7_cpumask;
+
 static bool domain_is_valid(unsigned domain)
 {
 	switch (domain) {

@@ -1641,6 +1643,44 @@ static struct pmu h_24x7_pmu = {
 	.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
 };
 
+static int ppc_hv_24x7_cpu_online(unsigned int cpu)
+{
+	/* Make this CPU the designated target for counter collection */

The comment implies every newly onlined CPU will become the target, but
actually it's only the first onlined CPU.

So I think the comment needs updating, or you could just drop the
comment, I think the code is fairly clear by itself.

Hi Michael,
   Thanks for reviewing the patch. Sure I will update it accordingly.

quoted

+	if (cpumask_empty(&hv_24x7_cpumask))
+		cpumask_set_cpu(cpu, &hv_24x7_cpumask);
+
+	return 0;
+}
+
+static int ppc_hv_24x7_cpu_offline(unsigned int cpu)
+{
+	int target = -1;

No need to initialise target, you assign to it unconditionally below.

Ok Will change.

quoted

+	/* Check if exiting cpu is used for collecting 24x7 events */
+	if (!cpumask_test_and_clear_cpu(cpu, &hv_24x7_cpumask))
+		return 0;
+
+	/* Find a new cpu to collect 24x7 events */
+	target = cpumask_last(cpu_active_mask);

Any reason to use cpumask_last() vs cpumask_first(), or a randomly
chosen CPU?

Incase we sequentially offline multiple cpus, taking cpumask_first() may add some latency in
that scenario.

So, I was trying to test benchmark in power9 lpar with 16 cpu, by off-lining cpu 0-14

With cpumask_last: This is what I got.

real	0m2.812s
user	0m0.002s
sys	0m0.003s

With cpulast_any:
real	0m3.690s
user	0m0.002s
sys	0m0.062s

That's why I just went with cpumask_last thing. Please Let me know if any changes required.

quoted

+	if (target < 0 || target >= nr_cpu_ids)
+		return -1;
+
+	/* Migrate 24x7 events to the new target */
+	cpumask_set_cpu(target, &hv_24x7_cpumask);
+	perf_pmu_migrate_context(&h_24x7_pmu, cpu, target);
+
+	return 0;
+}
+
+static int hv_24x7_cpu_hotplug_init(void)
+{
+	return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_24x7_ONLINE,
+			  "perf/powerpc/hv_24x7:online",
+			  ppc_hv_24x7_cpu_online,
+			  ppc_hv_24x7_cpu_offline);
+}
+
 static int hv_24x7_init(void)
 {
 	int r;

@@ -1685,6 +1725,11 @@ static int hv_24x7_init(void)
 	if (r)
 		return r;
 
+	/* init cpuhotplug */
+	r = hv_24x7_cpu_hotplug_init();
+	if (r)
+		pr_err("hv_24x7: CPU hotplug init failed\n");
+

The hotplug initialisation shouldn't fail unless something is badly
wrong. I think you should just fail initialisation of the entire PMU if
that happens, which will make the error handling in the next patch much
simpler.

I will update it.

Thanks,
Kajol Jain

cheers

quoted

 	r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
 	if (r)
 		return r;